Read the classification and you read the invention's center of gravity. Google's grant US10946515B2 ("Deep machine learning methods and apparatus for robotic grasping," issued March 16, 2021; inventors Sergey Levine, Peter Pastor Sampedro, Alex Krizhevsky) hangs on G06N 3/08 and 3/084 — neural-network training and backpropagation — as much as on the B25J control codes. The claim isn't about a gripper. It's about how a network learns to predict good grasps.
The mechanism is learned grasp prediction: a deep network ingests visual input and outputs grasp parameters, trained on large-scale grasp-attempt data. The B25J 9/1664 (control based on sensed object) and 9/1697 (vision-sensor control) codes describe the robot acting on the network's output; the G06N codes describe the learning that produces it. The inventive weight sits on the learning side, which is where Google's advantage has always been.
For the manipulation beat, this is the inverse of a mechanism patent. Where Mitsubishi's gripper claim fences finger geometry, Google's claim fences a training method — and the two age very differently. A learned-grasping claim is exposed to fast-moving prior art (the whole deep-RL-for-manipulation literature) but is also extremely broad in what hardware it can read on, because any gripper driven by a network of the claimed kind is in scope.
From a portfolio angle, this is canonical Google-robotics IP: protect the algorithm and the training apparatus, stay hardware-agnostic. The named inventors are central figures in learned manipulation, and the family (Google has continuations in this space) signals a sustained fence around grasp-learning rather than a single shot. That breadth-by-algorithm posture is exactly how a software company builds a manipulation moat.
Caveats, and they're sharp here. Learned-grasping is one of the most heavily-published areas in robotics; a broad training-method claim faces intense obviousness and prior-art exposure, and granted scope may hinge on a specific training detail. Software-method claims also age fast as architectures change. Read claim 1 for the exact training step — that limitation, not "deep learning for grasping," is the asset.
For the file: an algorithm-centric, hardware-agnostic grasp-learning grant from the field's leading lab. Pull US10946515B2 on PatentBear and read claim 1's training limitation — and check the continuations, because the family is where the real scope lives.